-
Notifications
You must be signed in to change notification settings - Fork 0
/
fixed_noise_dataset.py
142 lines (114 loc) · 5.92 KB
/
fixed_noise_dataset.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
import abc
import typing
import torch
import torch.utils.data
from deterministic_multivariate_normal import DeterministicMultivariateNormal
class Noise(abc.ABC):
def __init__(self) -> None:
super().__init__()
self._generator = torch.Generator()
self.__initial_generator_state = self._generator.get_state()
@abc.abstractmethod
def __call__(self, x: torch.Tensor) -> typing.Tuple[torch.Tensor, torch.Tensor]:
...
def reset(self, seed: typing.Union[int, None] = None) -> None:
self._generator.set_state(self.__initial_generator_state)
if seed is not None:
self._generator.manual_seed(seed)
def __getstate__(self) -> tuple[list[int], list[int]]:
return (self.__initial_generator_state.tolist(), self._generator.get_state().tolist())
def __setstate__(self, state: tuple[list[int], list[int]]) -> None:
self._generator = torch.Generator()
self.__initial_generator_state = torch.tensor(state[0], dtype=torch.uint8)
self._generator.set_state(torch.tensor(state[1], dtype=torch.uint8))
class AdditiveElementwiseUniformNoise(Noise):
def __init__(self, min_: float = 0.0, max_: float = 1.0) -> None:
super().__init__()
self.__min = min_
self.__max = max_
def __call__(self, x: torch.Tensor) -> typing.Tuple[torch.Tensor, torch.Tensor]:
noise = self.__min+(self.__max-self.__min)*torch.rand(x.shape, dtype=x.dtype, device=x.device, generator=self._generator)
return (x+noise, noise)
def __getstate__(self) -> tuple[tuple[list[int], list[int]], float, float]: # type: ignore
return (super().__getstate__(), self.__min, self.__max)
def __setstate__(self, state: tuple[tuple[list[int], list[int]], float, float]) -> None: # type: ignore
super().__setstate__(state[0])
self.__min = state[1]
self.__max = state[2]
class AdditiveElementwiseGaussianNoise(Noise):
def __init__(self, mu: float = 0.0, sigma: float = 1.0) -> None:
super().__init__()
self.__mu = mu
self.__sigma = sigma
def __call__(self, x: torch.Tensor) -> typing.Tuple[torch.Tensor, torch.Tensor]:
noise = self.__mu+self.__sigma*torch.randn(x.shape, dtype=x.dtype, device=x.device, generator=self._generator)
return (x+noise, noise)
def __getstate__(self) -> tuple[tuple[list[int], list[int]], float, float]: # type: ignore
return (super().__getstate__(), self.__mu, self.__sigma)
def __setstate__(self, state: tuple[tuple[list[int], list[int]], float, float]) -> None: # type: ignore
super().__setstate__(state[0])
self.__mu = state[1]
self.__sigma = state[2]
class AdditiveElementwisePoissonNoise(Noise):
def __init__(self, rate: float = 1.0) -> None:
super().__init__()
self.__rate = rate
def __call__(self, x: torch.Tensor) -> typing.Tuple[torch.Tensor, torch.Tensor]:
noise = torch.poisson(torch.full(x.shape, self.__rate, dtype=x.dtype, device=x.device), generator=self._generator)
return (x+noise, noise)
def __getstate__(self) -> tuple[tuple[list[int], list[int]], float]: # type: ignore
return (super().__getstate__(), self.__rate)
def __setstate__(self, state: tuple[tuple[list[int], list[int]], float]) -> None: # type: ignore
super().__setstate__(state[0])
self.__rate = state[1]
class AdditiveTensorwiseGaussianNoise(Noise):
def __init__(self, mu: torch.Tensor, sigma: typing.Union[torch.Tensor, None] = None) -> None:
super().__init__()
if sigma is None:
sigma = torch.diag_embed(torch.ones((mu.numel())))
self.__distribution = DeterministicMultivariateNormal(mu, sigma)
def __call__(self, x: torch.Tensor) -> typing.Tuple[torch.Tensor, torch.Tensor]:
noise = self.__distribution.rsample(x.shape, generator=self._generator).to(x.dtype).to(x.device)
return (x+noise, noise)
def __getstate__(self) -> tuple[tuple[list[int], list[int]], float, float]: # type: ignore
raise NotImplementedError()
def __setstate__(self, state: tuple[tuple[list[int], list[int]], float, float]) -> None: # type: ignore
raise NotImplementedError()
class FixedNoiseDataset(torch.utils.data.Dataset[typing.Tuple[torch.Tensor, ...]]):
def __init__(self,
dataset: torch.utils.data.Dataset[typing.Tuple[torch.Tensor, ...]],
noisy_features: typing.Sequence[int] = (0,),
noise: typing.Union[Noise, None] = None,
append_clean: bool = False,
append_noise: bool = False) -> None:
super().__init__()
self.__dataset = dataset
self.__noisy_features = noisy_features
if noise is None:
noise = AdditiveElementwiseGaussianNoise()
self.__noise = noise
self.__seeds = torch.randint(10000000,99999999, (len(typing.cast(typing.Sized, dataset)),len(noisy_features)))
self.__append_clean = append_clean
self.__append_noise = append_noise
def __len__(self) -> int:
return len(typing.cast(typing.Sized, self.__dataset))
def __getitem__(self, i: int) -> typing.Tuple[torch.Tensor, ...]:
sample = self.__dataset[i]
transformed_features = []
clean_features = []
noise_list = []
for j in range(len(sample)):
if j in self.__noisy_features:
self.__noise.reset(int(self.__seeds[i,j].item()))
noisy_feature, noise = self.__noise(sample[j])
transformed_features.append(noisy_feature)
clean_features.append(sample[j])
noise_list.append(noise)
else:
transformed_features.append(sample[j])
result = [*transformed_features]
if self.__append_clean:
result += clean_features
if self.__append_noise:
result += noise_list
return tuple(result)